Silero TTS

The Silero Team's Russian speech synthesizer, and the reason this spoke's account of the tts-front-end needs amending. 6.1k★ on snakers4/silero-models, which ships the TTS models alongside Silero's STT and VAD.

The spoke already held silero-stress — the same team's standalone stress placer, MIT — without holding the synthesizer it was built for. This page closes that.

The Russian line

Five speakers throughout: aidar, baya, kseniya, xenia, eugene (v4_ru also offers random). Sample rates 8000 / 24000 / 48000 Hz.

V5 models support SSML. Auto-stress and homograph resolution are Russian-only — the multilingual and Indic lines get neither.

The front-end came inside the model

This is the fact that changes something. synthesis.md has said since 2026-08-27 that "a synthesizer supporting Russian either has a stress model in front of it or is guessing," with ruaccent as the worked example of what that stage costs when built separately: five ONNX models, 21 MB of dictionaries, three years of one person's work.

Silero v5 absorbed the stage. Stress placement, ё restoration and homograph disambiguation are flags on the synthesis call — put_accent, put_yo, put_stress_homo, put_yo_homo, not a separate library. The team ships silero-stress standalone under MIT and runs the same job inside the synthesizer under a different licence.

So there are now three shapes for the same stage in this corpus: bolt-on (ruaccent), standalone library from the model vendor (silero-stress), and folded into synthesis (here). The synthesis claim survives — a Russian synthesizer still needs a stress model — but "in front of it" was too narrow.

Speed, as the vendor measured it

From Silero's own release write-up on Habr (published 2025-10-31, T3 — the maker reporting on the maker), on an Intel Core i9-10940X and an RTX 3090:

Claimed 3–4× faster than v3 and 1.5–2× faster than v4. Model size ~140 MB, described as compressed without much optimization effort. Optimal input length is stated as 5–20 seconds of audio on CPU, 10+ on GPU.

Nobody outside Silero has checked any of this. russian-tts-evaluation measured v3_1 and v4 on independent hardware and stopped there, a year before v5 existed.

What the independent numbers said about v4

russian-tts-evaluation (2024-07-12) is unflattering and worth keeping next to the vendor table:

• v3_1 took the joint-best clarity in the field — CER 0.7, beaten only by Yandex at 0.6.
• v4 was worse than v3_1 on both quality axes: CER 1.0 against 0.7, and UTMOS 1.755 against 2.544, the lowest studio-quality score of the twenty-one rows tested. It bought that with speed (CPU xRT 0.0544 against 0.1256).
• Intonation is where it loses badly. Encodec FAD 79–170 against Vosk-TTS's 9.99. The evaluator's own summary, plain intonation and hard-to-beat clarity, describes Silero exactly.

Two versions have shipped since, both claiming quality gains, neither measured by anyone else.

The licence, which decides most audiobook questions

CC BY-NC-SA for the TTS models, including every v4_ru and v5_ru variant above. The exception is narrow and easy to misread: only the base CIS models are MIT.

Non-commercial. A narrated audiobook you intend to sell is not covered, and neither is one behind a paywall or carrying ads. This puts Silero on the same side of open-weight-tts's licence axis as voicecraft (CC BY-NC-SA + CPML) and XTTS-v2 (CPML) — and on the opposite side from kokoro and orpheus. Note the contrast inside one team's own output: silero-stress, which does the hard linguistic half, is MIT; the synthesizer is not.

Questions, and a caveat on the date

v5_4_ru added interrogative intonation across four question types: special, general, alternative, and tag questions. General questions need the intonational centre marked with asterisks, because which word carries the accent changes what the sentence means. An intensity=3 parameter controls how strongly the contour is applied; the team recommends leaving defaults alone.

The write-up for this feature returned a publication date of 2024-03-27 through our fetch, which cannot be right for a v5 model when v5 itself shipped in October 2025. The date is not recorded on this page, per the rule against guessing one. The feature claims themselves are on the repo's own model table and stand independently.

Related

russian-tts · silero-stress · ruaccent · tts-front-end · russian-tts-evaluation · open-weight-tts · text-to-speech · vosk